HanaFEKI/Silver_Medal-Clinical_reasoning_challenge

NLP solution for the Zindi challenge predicting clinician responses to real Kenyan healthcare vignettes. Fine-tuned a bart-large model under strict resource limits, achieving a Silver medal (rank 61/440). Focused on clinical reasoning with limited data in low-resource settings. # 🥈 Zindi Challenge: Kenya Clinical Reasoning Challenge – Silver Medal This repository contains my solution to the **Kenya Clinical Reasoning Challenge** hosted on the Zindi platform. I ranked **61st out of 440 participants**, earning a **Silver Medal** 🥈 by building an NLP-based model that attempts to emulate the clinical reasoning of frontline Kenyan healthcare workers. ## 🩺 Context In resource-limited settings like rural Kenya, nurses must make high-stakes medical decisions with limited tools and specialist backup. This challenge provided **400 authentic clinical vignettes**—each simulating a real-world scenario involving patient presentation, facility type, and nurse background. The objective was to **predict the most appropriate clinician response** to each case. Each response was previously evaluated by human experts and compared with top-tier AI models including **GPT-4**, **Gemini**, and **LLaMA**. ## 📦 About the Challenge The dataset comprises **authentic clinical vignettes** sourced from Kenyan frontline healthcare environments. Each vignette describes a case scenario with relevant context such as: - Patient symptoms and background - Nurse experience level - Facility type (e.g., dispensary, hospital) Our task was to **predict the clinician’s written response** to each scenario, responses may include medical abbreviations, structured reasoning (e.g., "Summary:", "Diagnosis:", "Plan:"), or free-form clinical notes. > 🔍 These vignettes simulate daily decisions made by nurses under resource constraints, often without access to advanced diagnostics or specialist support. ## 📊 Dataset Summary - `train.csv`: 400 prompts and clinician responses - `test.csv`: 100 prompts (unlabeled) - Data is small but **high quality**, curated by domain experts - Tasks span **multiple medical specialties**, **geographies**, and **clinical contexts** - Responses are free text, requiring **robust natural language understanding** ## 🧠 My Approach ### 🔹 1. Text P …